Exploiting semantics for shape-based 3D model retrieval

Abstract

A shape similarity judgment among a pair of 3D models is often influenced by their semantics, in addition to their shapes. In this paper, we present a method to improve shape-based 3D model retrieval performance by learning multiple semantic categories off-line from a small set of training examples. Learning multiple semantic categories at a time from a small number of labeled training samples whose features have high-dimensionality has been quite difficult. In our proposed method, we apply unsupervised learning to partially purify the set of features so that their saliency improves. Then, using a supervised learning algorithm captures the set of semantic categories from the partially purified feature set. Our experimental evaluation showed that the retrieval performance using the proposed method is significantly higher than those of both supervised-only and unsupervised-only learning methods.